CONTINUOUS BLOOD PRESSURE ESTIMATION USING TRADITIONAL ML AND DL TECHNIQUES

Abstract

The project aims to enable continuous, non-invasive blood pressure (BP) estimation using physiological signals—PPG, ECG, and PCG—through a combination of deep learning and traditional machine learning models. Motivated by the need for real-time cardiovascular monitoring, the work addresses a critical healthcare challenge with potential applications in smartwatches and edge devices. The proposed solution follows a dual-track approach: one using raw signal data, and another using feature-engineered data. For the raw data track, various preprocessing techniques—including butterworth filtering, moving average smoothing, and discrete wavelet transform—were evaluated, with the best results achieved by an advanced LSTM model trained on PPG+ECG signals filtered with butterworth+MAF (4.03 mmHg for SBP, 2.27 mmHg for DBP). In the feature-engineered track, statistical and physiological features were extracted, and models like RF, GBR, LSTM, and a Transformer inspired by the TransfoRhythm architecture were tested, with the RF Regressor delivering the best performance (MAE of 1.2 mmHg for SBP and 0.3 mmHg for DBP). The project also explored ABP prediction from raw PCG signals on basic CNN, Transformer and LSTM-based models, extending the scope to multimodal signal fusion. CNN achieved the lowest MAE of 13.19 mmHg.